Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

时间：2021-05-19 15:57:05

Tuning process

下图中的需要tune的parameter的先后顺序， 红色>黄色>紫色，其他基本不会tune.

先讲到怎么选hyperparameter, 需要随机选取(sampling at random)

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks　　

随机选取的过程中，可以采用从粗到细的方法逐步确定参数

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

有些参数可以按照线性随机选取， 比如 n[l]

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

但是有些参数就不适合线性的sampling at radom, 比如 learning rate α，这时可以用 log

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

Andrew 很幽默的讲到了两种选参数的实际场景 pandas vs caviar. pandas approach 一般用在你的算力不够时候，要持续几天的training.

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

Batch norm

我们知道对input layer 做 normalizing, 其实对每一层的输入都可以做normalizing, 这就是 batch norm. 做batch norm 时，有对 activation后的结果做norm 的，也有对activation 前的结果 z 做batch norm 的，这里讲的是后一种，对z 做norm.

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

为什么Batch Norm 起作用呢？

先看下下面图讲到的convariate shift，如果traing set 的distribution 变了，就应该重新train model. 同样，对NN的每一层也有类似的问题.

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

Andrew讲到batch norm 是为了尽量使得不同layer decouple，这样相互影响就要小一点，整个NN比较稳定.

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

Batch norm 还有regularization 的作用，但是这个算法主要不是做这个的. 不建议专门用它来做regularization.

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

对 test set 求 μ, σ2， 采用了不一样的方法，就是基于签名mini-batch set 求出来的μ, σ2 应用exponetially weighted average 求平均值. 它和logistic regression 一样，decision boudary 是线性的.

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

Softmax Regression

Softmax regression 就是 logistic regression 的generaliazation 版本， 它可以用在multi-class clarification 问题上。和logistic regression 一样，decision boudary 都是线性的. 如果要使得decison boudary 是非线性的就需要deep network.

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

Programing framework

TensorFlow by google, an example

　　Coursera Deep Learning 2 Improving Deep Neural Networks: Hyperparameter tuning, Regularization and Optimization - week3, Hyperparameter tuning, Batch Normalization and Programming Frameworks

相关文章